The name 'probe' (and thus 'probe group') is a relic from the times when
sigrok was mostly about logic analyzers. Nowadays we support a lot more
device types where 'probe' is not really a good term and 'channel' is
much better suited.
This fixes parts of bug #259.
Instead of >= 44 Makefile.am's we now only have one top-level
Makefile.am, and use the 'subdir-objects' automake option to
handle the build via non-recursive (auto)make.
This has the advantage of fewer (boilerplate or other) files and less
clutter in general, as well as performance advantages since the new
setup can build many files in parallel (with 'make -j'), not only 2 or 3
files within the same (e.g. hardware/xxxx/* subdirectory) and also since
we no longer need to build intermediate libtool helper libs per subdirectory.
A quick, non-scientific test build on a quad-core laptop with 'make -j 4'
yields a build time reduction from 35s to 19s.
All autotools features that worked before are still intact without any
regressions, including the Make targets 'install', 'uninstall', 'check',
'dist', 'clean', 'distclean' and so on, as well as all the usual portability
handling (build works on any OS, with any Make implementation such as
GNU Make or BSD Make, with any shell such as sh/ksh/zsh/bash/dash, etc. etc.)
and features such as out-of-tree build support, cross-compile support,
testsuite support (also with colored output), "silent make rules", etc. etc.
Implement the configuration setting TRIGGER_SOURCE with the
choices CH (logic channels) and TRG (external trigger input).
Also implement the TRIGGER_SLOPE setting for selecting the
edge to trigger on (rising or falling).
It turns out that all LWLA protocol responses consist either
of 32-bit units or of 32-bit units combined into 64-bit units.
Thus it makes sense to double the basic unit size for reading
from 16 bit to 32 bit.
We cannot do the same for command messages though, as those
actually do use 16-bit quantities in some places, and 32-bit
arguments are not always aligned to 32-bit boundaries.
(acquisition_state.xfer_buf_in): Change unit type to uint32_t,
and update related macros and code accordingly.
(LWLA_TO_UINT32): New macro to replace LWLA_READ32, operating
directly on 32-bit values instead of pointers to 16-bit units.
Make use of a compiler-recognized idiom for bitwise rotation
to efficiently swap the 16-bit halves of a 32-bit word.
(LWLA_TO_UINT16): New macro to replace LWLA_READ16.
(LWLA_READ64): Remove unused macro.
(LWLA_WORD_[0123]): Slightly simplify 16-bit word extraction.
The return code SR_ERR_ARG is intended for reporting unsupported
or inapplicable device configuration settings and is not a hard
error. In order to indicate failure of internal sanity checks,
use SR_ERR_BUG instead.
Modify the bitstream loading routine to work directly with the
Raw Binary Files (.rbf) generated by Altera tools. Previously,
a custom format was used which was basically an RBF preceded by
a 4-byte header specifying the transfer length.
Move pre-acquisition hardware setup to the new config_commit()
callback. At the moment, the only setting applied at commit
time is switching the clock source, which involves uploading
a new bitstream to the FPGA.
Move setup of channels and trigger masks to the new probe
configuration callback. Although the actual hardware setup
still happens just before acquisition, the new approach
already has the advantage that invalid settings are caught
early.
Also, it turns out that the LWLA1034 allows triggering on
channels which are not enabled for data acquisition. This
feature is now supported as well.
Apparently, frontends may call scan() more than once to accumulate
multiple devices, so do not reset the instance list pointer at the
start of each scan. Also, number devices continuously across scans.
Report settings at acquisition start as informational messages.
Print a message when the the trigger condition has been met.
Demote some other messages from information to debug, and use
the %zu format for printing size_t values.
(process_sample_data): When expanding run-length samples into
session packets, calculate the number of samples to write in
advance while honoring all constraints. This is cleaner than
checking constraints within the expansion loop. Also, the new
logic always fills up packets exactly to whatever limit applies
first, thereby removing the need for truncation after the fact.
Allow the acquisition to be constrained by time in addition to
a sample count limit. Since the LWLA protocol actually provides
only a duration natively, implement the sample count limit on top
of the new duration limit.
With this change, limiting an acquisition in external clock mode
should finally work properly.
(lwla_setup_acquisition): Set the clock divider bypass
flag to 1 for the external clock modes as well.
(capture_setup): Set the clock divide count to 0 if an
external clock source is selected.
(lwla_send_bitstream): Unref the mapped file earlier in order
to simplify the error handling.
(lwla_receive_reply): Do not treat a reply buffer length of
zero as silent no-op. That logic was left over from an earlier
iteration, before the distinction between reply buffer size and
expected read length was introduced.
(lwla_set_clock_source): Checking whether an enum value is greater
than or equal to zero apparently results in a warning with some
compilers. Assign the enum to an unsigned variable to avoid this,
and return SR_ERR_BUG if the range is exceeded, as this indicates
a bug in the driver code itself.
The g_ascii_formatd() function expects the "format" argument to start
with a '%' character, e.g. it should be "%f" or such (this is not
clearly documented in the glib API docs, but visible from the source code).
The usage of "CH%f" for example will trigger an assertion and thus make the
LWLA device unusable in practice (e.g. in PulseView on Windows no probenames
would be shown, and sampling wouldn't work).
Example:
GLib-CRITICAL **: g_ascii_formatd: assertion 'format[0] == '%'' failed
(not exposed in all glib versions or builds of glib on all distros
apparently, some may need G_MESSAGES_DEBUG=all or other measures)
From the glib g_ascii_formatd() code:
g_return_val_if_fail (format[0] == '%', NULL);
We now use g_snprintf() instead for simplicity. This has been tested to
fix this specific issue (i.e. the probenames now do show up in PulseView).
This closes bug #270.
Uwe Hermann
Daniel Elstner